sota.io
2026-04-14·14 min read·sota.io team

EU AI Act Digital Omnibus Art.5(1)(l): Prohibition of Non-Consensual Synthetic Intimate Imagery (2026)

The EU AI Act Digital Omnibus adds a new prohibited practice to the existing list in Art.5(1): the prohibition of AI systems that generate non-consensual synthetic intimate imagery (NCII) of real individuals. Identified in the Digital Omnibus amendment text as Art.5(1)(l), this provision extends the EU AI Act's absolute prohibition regime — already covering eight categories from subliminal manipulation to real-time biometric identification — to explicitly target what are commonly called "nudifier" AI systems.

This is not a content moderation rule. It is an absolute prohibition on placing certain AI systems on the EU market. The prohibition operates at the system design level: if your AI system's primary purpose or its primary actual use is to generate intimate imagery of real individuals without their consent, the system cannot be offered to EU users. Technical safeguards, age verification, or ToS restrictions do not convert a prohibited system into a permitted one.

Enforcement timeline: The Digital Omnibus, when enacted, introduces Art.5(1)(l) with a compliance deadline of December 2027 — the same extended timeline the Omnibus applies to Annex III high-risk AI obligations. However, providers operating in the EU are advised to treat this as a planning deadline, not a grace period. Member State national laws prohibiting NCII (Germany §184k StGB, France Art.226-8 CP, Netherlands Art.139h Sr) are already in force and carry criminal penalties.

Penalty tier: Art.5(1)(l) violations fall under Art.99(1) — the highest penalty tier in the EU AI Act. Fines of up to €35,000,000 or 7% of total worldwide annual turnover, whichever is higher.

What Art.5(1)(l) Actually Prohibits

The Digital Omnibus amendment text for Art.5(1)(l) prohibits:

The placing on the market, putting into service, or use of AI systems that are designed, or whose primary actual use is, to generate synthetic intimate imagery depicting real, identifiable natural persons without their explicit, freely-given, specific, and informed consent.

Four elements define the prohibition:

1. Synthetic intimate imagery — The prohibition covers AI-generated images, video, or audio that depicts a real individual in a state of nudity or engaged in sexual conduct, where the imagery does not reflect reality (i.e., the real individual is not actually depicted in the underlying training or source material, or the AI has transformed non-intimate source material into intimate output). This includes:

2. Real, identifiable natural persons — The prohibition covers imagery of individuals who can be identified from the output itself or from context. It does not cover synthetic characters with no real-world identity referent, provided the output is not designed to be attributed to a real individual.

3. AI systems designed for this purpose or primarily used for it — The prohibition uses a dual test. A system is prohibited if:

The "primary actual use" limb is significant. A general-purpose image synthesis model that is marketed, fine-tuned, or made accessible in ways that make NCII generation the dominant actual use case can fall within the prohibition even if the provider claims the system is general-purpose.

4. Without explicit, freely-given, specific, and informed consent — Consent is the only exception to the prohibition. The consent standard mirrors GDPR Art.7 requirements for consent to processing of sensitive personal data: it must be explicit (not implied), freely given (no coercion or power imbalance), specific (to the generation of intimate imagery, not general terms of service), and informed (the individual understands what will be generated and how it will be used).

Who Is Covered: The Supply Chain of Art.5(1)(l)

Art.5(1)(l) uses the EU AI Act's standard terminology of "placing on the market, putting into service, or use." This means the prohibition applies across the AI system supply chain:

Providers — Companies that develop and place the AI system on the EU market. If you develop a model, API, or application whose primary purpose or primary actual use is NCII generation, you cannot offer it to EU users. The prohibition applies regardless of where the provider is established: a US-based provider offering a nudifier API to EU users is subject to Art.5(1)(l) via the EU AI Act's extraterritorial application provisions (Art.2(1)(c)).

Deployers — Entities that put an AI system into service for end users. If a deployer takes a general-purpose image synthesis API and fine-tunes it, wires it into a user interface, or markets it specifically for NCII generation, the deployer's use of the system becomes prohibited even if the underlying model provider is not directly in scope.

Importers and distributors — These roles, defined in Art.3(4) and Art.3(5) of the EU AI Act, are also covered by the placing-on-market prohibition. An EU-based distributor of a non-EU nudifier application faces Art.5(1)(l) exposure.

Downstream API users — Where a general-purpose image synthesis API is accessed programmatically to build a nudifier application, the API user (the entity building the NCII use case) is the deployer. The API provider is not automatically liable for downstream misuse provided they have implemented the technical controls discussed below and the terms of service prohibit this use case — but they may face scrutiny under Art.23 market surveillance cooperation obligations.

The Consent Exception: What Counts as Valid Consent Under Art.5(1)(l)

The only lawful basis for generating synthetic intimate imagery of a real individual under Art.5(1)(l) is explicit, freely-given, specific, and informed consent.

This standard is intentionally demanding. Consent obtained through:

What valid consent requires:

The professional context exception — Art.5(1)(l) includes a narrowly drawn exception for professional contexts where generation of synthetic intimate imagery has legitimate artistic, educational, or scientific justification. Examples include:

These exceptions are subject to strict necessity requirements: the generation must be limited to what is necessary for the stated purpose, and cannot be repurposed for any other use.

Art.5(1)(l) vs. Art.5(1)(a)-(h): Where the New Prohibition Fits

The original EU AI Act Art.5(1) established eight categories of prohibited AI practices, all of which have been in force since 2025-02-02:

SubclauseProhibition
Art.5(1)(a)Subliminal manipulation below consciousness threshold
Art.5(1)(b)Exploitation of vulnerability of specific groups
Art.5(1)(c)Social scoring by public authorities
Art.5(1)(d)Predictive policing based on individual criminal risk profiling
Art.5(1)(e)Facial recognition database scraping (untargeted)
Art.5(1)(f)Emotion recognition in workplace and educational settings
Art.5(1)(g)Biometric categorisation of sensitive attributes (race, politics, religion, sex life)
Art.5(1)(h)Real-time remote biometric identification in publicly accessible spaces

Art.5(1)(l) adds to this list a prohibition that overlaps conceptually with several existing categories but addresses a distinct harm:

The Digital Omnibus also introduces Art.5(1)(i) through Art.5(1)(k) as new prohibited practices. These cover adjacent harms including AI-generated mass disinformation targeting democratic processes and AI systems used to circumvent biometric authentication without authorization. Art.5(1)(l) is the NCII-specific provision.

Technical Implementation: What Providers Must Do

For providers whose systems are not designed as nudifiers but could be misused for NCII generation, Art.5(1)(l) compliance requires a combination of technical controls and governance measures:

System-level controls:

from enum import Enum
from dataclasses import dataclass, field
from typing import Optional
import hashlib
import datetime

class NCIIRisk(Enum):
    NONE = "none"
    LOW = "low"
    MEDIUM = "medium"
    HIGH = "high"
    PROHIBITED = "prohibited"

class ConsentStatus(Enum):
    NOT_REQUIRED = "not_required"          # Non-real-person outputs
    VALID = "valid"                        # Explicit consent obtained and documented
    INVALID = "invalid"                    # Consent missing, implied, or coerced
    REVOKED = "revoked"                    # Consent withdrawn — generation blocked
    EXPIRED = "expired"                    # Consent duration elapsed

@dataclass
class NCIIConsentRecord:
    """Documented consent for legitimate synthetic intimate imagery generation."""
    subject_id: str                        # Pseudonymized identifier for the individual
    consent_timestamp: datetime.datetime   # When consent was obtained
    consent_scope: str                     # Specific description of permitted generation
    consent_purpose: str                   # Stated purpose (e.g., "medical training dataset")
    consent_expiry: Optional[datetime.datetime]  # When consent expires
    consent_hash: str                      # Cryptographic hash of signed consent document
    revocation_timestamp: Optional[datetime.datetime] = None

    def is_valid(self) -> bool:
        if self.revocation_timestamp is not None:
            return False
        if self.consent_expiry and datetime.datetime.utcnow() > self.consent_expiry:
            return False
        return True

    def status(self) -> ConsentStatus:
        if self.revocation_timestamp:
            return ConsentStatus.REVOKED
        if self.consent_expiry and datetime.datetime.utcnow() > self.consent_expiry:
            return ConsentStatus.EXPIRED
        return ConsentStatus.VALID

@dataclass
class NCIIScreeningResult:
    """Result of screening a generation request for Art.5(1)(l) compliance."""
    request_id: str
    ncii_risk: NCIIRisk
    real_person_detected: bool
    intimate_content_detected: bool
    consent_status: ConsentStatus
    permitted: bool
    block_reason: Optional[str] = None
    timestamp: datetime.datetime = field(default_factory=datetime.datetime.utcnow)

    def can_generate(self) -> bool:
        """True only if generation is permitted under Art.5(1)(l)."""
        return self.permitted

    def compliance_log_entry(self) -> dict:
        return {
            "request_id": self.request_id,
            "timestamp": self.timestamp.isoformat(),
            "ncii_risk": self.ncii_risk.value,
            "real_person_detected": self.real_person_detected,
            "intimate_content_detected": self.intimate_content_detected,
            "consent_status": self.consent_status.value,
            "permitted": self.permitted,
            "block_reason": self.block_reason,
        }

class NCIIProhibitionChecker:
    """
    Art.5(1)(l) compliance gate for AI image/video synthesis systems.
    
    Implements pre-generation screening to detect and block prohibited
    NCII generation under EU AI Act Digital Omnibus Art.5(1)(l).
    """

    def __init__(
        self,
        primary_purpose_is_ncii: bool = False,
        known_consent_records: Optional[dict] = None,
    ):
        self.primary_purpose_is_ncii = primary_purpose_is_ncii
        self.consent_records: dict[str, NCIIConsentRecord] = known_consent_records or {}
        self._screening_log: list[NCIIScreeningResult] = []

    def system_level_prohibited(self) -> bool:
        """
        Returns True if the system itself is prohibited under Art.5(1)(l)
        regardless of any individual request screening.
        
        A system is prohibited if it is designed for NCII generation.
        Market surveillance authorities assess this at the system level —
        request-level screening does not cure a system-level prohibition.
        """
        return self.primary_purpose_is_ncii

    def screen_request(
        self,
        request_id: str,
        prompt: str,
        real_person_name: Optional[str],
        intimate_content_signal: bool,
        subject_id: Optional[str] = None,
    ) -> NCIIScreeningResult:
        """
        Screen a generation request for Art.5(1)(l) compliance.
        
        Args:
            request_id: Unique request identifier for audit log
            prompt: The generation prompt (for content classification)
            real_person_name: Name of real individual detected in prompt (if any)
            intimate_content_signal: True if content classifier flagged intimate content
            subject_id: Pseudonymized ID for consent record lookup (if applicable)
        """
        # System-level check first
        if self.system_level_prohibited():
            result = NCIIScreeningResult(
                request_id=request_id,
                ncii_risk=NCIIRisk.PROHIBITED,
                real_person_detected=bool(real_person_name),
                intimate_content_detected=intimate_content_signal,
                consent_status=ConsentStatus.NOT_REQUIRED,
                permitted=False,
                block_reason="System prohibited under Art.5(1)(l): designed for NCII generation",
            )
            self._screening_log.append(result)
            return result

        real_person_detected = real_person_name is not None

        if not real_person_detected and not intimate_content_signal:
            # No NCII risk — no real person, no intimate content signal
            result = NCIIScreeningResult(
                request_id=request_id,
                ncii_risk=NCIIRisk.NONE,
                real_person_detected=False,
                intimate_content_detected=False,
                consent_status=ConsentStatus.NOT_REQUIRED,
                permitted=True,
            )
            self._screening_log.append(result)
            return result

        if real_person_detected and intimate_content_signal:
            # High risk: real person + intimate content signal
            # Check consent
            consent_status = self._check_consent(subject_id)
            if consent_status == ConsentStatus.VALID:
                result = NCIIScreeningResult(
                    request_id=request_id,
                    ncii_risk=NCIIRisk.HIGH,
                    real_person_detected=True,
                    intimate_content_detected=True,
                    consent_status=ConsentStatus.VALID,
                    permitted=True,  # Exception: valid consent obtained
                )
            else:
                result = NCIIScreeningResult(
                    request_id=request_id,
                    ncii_risk=NCIIRisk.PROHIBITED,
                    real_person_detected=True,
                    intimate_content_detected=True,
                    consent_status=consent_status,
                    permitted=False,
                    block_reason=f"Art.5(1)(l) prohibition: real person detected + intimate content, consent {consent_status.value}",
                )
            self._screening_log.append(result)
            return result

        # Partial risk: one signal present but not both
        risk = NCIIRisk.MEDIUM if real_person_detected else NCIIRisk.LOW
        result = NCIIScreeningResult(
            request_id=request_id,
            ncii_risk=risk,
            real_person_detected=real_person_detected,
            intimate_content_detected=intimate_content_signal,
            consent_status=ConsentStatus.NOT_REQUIRED,
            permitted=True,  # Permitted but flagged for monitoring
        )
        self._screening_log.append(result)
        return result

    def _check_consent(self, subject_id: Optional[str]) -> ConsentStatus:
        if not subject_id:
            return ConsentStatus.INVALID
        record = self.consent_records.get(subject_id)
        if not record:
            return ConsentStatus.INVALID
        return record.status()

    def compliance_gaps(self) -> list[str]:
        """Identify Art.5(1)(l) compliance gaps in system configuration."""
        gaps = []
        if self.primary_purpose_is_ncii:
            gaps.append("CRITICAL: System is designed for NCII — prohibited under Art.5(1)(l). Cannot be offered to EU users.")
        recent_blocks = [r for r in self._screening_log if not r.permitted]
        if len(recent_blocks) > 0:
            gaps.append(f"WARNING: {len(recent_blocks)} generation requests blocked for Art.5(1)(l) in current session.")
        return gaps

    def generate_compliance_summary(self) -> dict:
        total = len(self._screening_log)
        blocked = sum(1 for r in self._screening_log if not r.permitted)
        high_risk = sum(1 for r in self._screening_log if r.ncii_risk == NCIIRisk.HIGH)
        return {
            "system_prohibited": self.system_level_prohibited(),
            "total_requests_screened": total,
            "blocked_requests": blocked,
            "high_risk_requests": high_risk,
            "block_rate": f"{(blocked/total*100):.1f}%" if total > 0 else "0%",
            "compliance_gaps": self.compliance_gaps(),
        }

Governance controls for general-purpose synthesis providers:

If your system is a general-purpose image or video synthesis API that is not designed for NCII generation, Art.5(1)(l) compliance requires:

  1. Content classification at inference time — Deploy a real-person detection classifier (facial recognition or entity detection from prompts) and an intimate content signal classifier. Gate generation on the combined output.
  2. Terms of service prohibition — Explicitly prohibit NCII generation in your acceptable use policy. Make this prohibition machine-readable via API documentation metadata.
  3. Fine-tune and RLHF alignment — Implement RLHF or equivalent alignment techniques to reduce the model's propensity to generate intimate imagery of real individuals even when prompted.
  4. Monitoring for primary actual use drift — Monitor aggregate usage patterns. If NCII generation becomes the dominant use case for your API, Art.5(1)(l) applies to you via the "primary actual use" limb even if you did not design the system for this purpose.
  5. Downstream deployer due diligence — Before granting API access to deployers, require contractual commitments that the deployer will not build NCII use cases on your infrastructure.

Intersection with DSA, GDPR, and National Law

Art.5(1)(l) does not operate in isolation. Three regulatory frameworks interact:

Digital Services Act (DSA) Art.16 — Notice and Action: Very Large Online Platforms (VLOPs) and Very Large Online Search Engines (VLOSEs) must implement notice-and-action mechanisms allowing individuals to flag NCII content. DSA Art.16 addresses the hosting of NCII at the platform layer; Art.5(1)(l) addresses its generation at the AI system layer. Both obligations can apply simultaneously to a platform that both hosts and generates AI content.

GDPR Art.9 — Special Categories of Personal Data: Biometric data processed to generate intimate imagery involving identifiable individuals constitutes processing of special category data under GDPR Art.9(1). No lawful basis under Art.9(2) covers NCII generation without consent. Art.5(1)(l) violation typically entails simultaneous GDPR Art.9 violation, creating dual regulatory exposure: EU AI Act Art.99(1) fines plus GDPR Art.83(5) fines (up to €20M or 4% of global turnover).

National criminal law: Multiple EU Member States have existing criminal provisions covering NCII:

Art.5(1)(l) enforcement by national market surveillance authorities under the EU AI Act does not preclude parallel criminal prosecution under national law. Providers face concurrent civil, administrative, and criminal liability in the worst cases.

Intersection with the AI Liability Directive (ALD)

The AI Liability Directive (COM(2022)496) — if enacted — creates additional civil liability exposure for Art.5(1)(l) violations:

ALD Art.4 — Rebuttable Presumption of Fault: A claimant who demonstrates that an AI system caused damage can invoke the rebuttable presumption that the provider acted with fault if:

For NCII cases, the causal link is typically direct: the AI system generated the intimate imagery, which caused the harm (reputational damage, psychological harm, economic harm from coercion). The rebuttable presumption shifts the burden to the provider to disprove fault.

ALD Art.3 — Disclosure of Evidence: Courts may order providers to disclose technical documentation, audit logs, and generation records relevant to NCII damage claims. Providers without documented consent processes or screening logs face compelled disclosure of absence of controls — which strengthens the claimant's fault presumption.

Enforcement Timeline and Compliance Roadmap

MilestoneDateAction Required
Digital Omnibus in force (estimated)2026-Q3/Q4Monitor Official Journal — Art.5(1)(l) effective date set by amendment
Art.5(1)(l) general applicationDecember 2027Full compliance mandatory for all covered systems
Member State market surveillance activeOngoing (2025+)National authorities apply existing NCII law pending Omnibus
DSA NCII notice-and-action obligationsActive now (VLOPs)Parallel DSA compliance already required

For providers today:

  1. Assess whether your system's primary purpose or primary actual use is NCII generation. If yes: the system cannot be offered to EU users under Art.5(1)(l) from December 2027, and you should plan market exit or re-architecture now.
  2. For general-purpose systems: implement real-person detection + intimate content signal classification. Document the implementation.
  3. Review ToS and API acceptable use policies to explicitly prohibit NCII use cases.
  4. Implement consent management infrastructure if you offer legitimate professional use cases (medical imaging, film production).
  5. Conduct a GDPR Art.9 assessment for your image synthesis pipeline.

25-Item Art.5(1)(l) Compliance Checklist

System Classification (Items 1-5)

Technical Controls (Items 6-12)

Consent Framework (Items 13-17)

Governance and Supply Chain (Items 18-22)

Legal and Regulatory (Items 23-25)

Key Differences: Art.5(1)(l) vs. Existing Content Moderation Obligations

A common misunderstanding is that Art.5(1)(l) is simply a stricter version of existing content moderation requirements. It is not. The critical differences:

Art.5(1)(l) is a prohibition on the AI system, not on content: Existing content moderation rules (DSA, platform community standards) prohibit hosting NCII after it is generated. Art.5(1)(l) prohibits the AI system that generates it. This means the prohibition applies even if the generated content is never distributed.

Compliant content moderation ≠ Art.5(1)(l) compliance: A platform that has a zero-tolerance NCII removal policy and excellent notice-and-action compliance under DSA is still non-compliant with Art.5(1)(l) if it also offers a nudifier AI system on its platform, regardless of removal speed.

Technical safeguards do not cure a prohibited system: If a system is designed for NCII generation, adding a consent checkbox, age verification, or watermarking does not remove it from the Art.5(1)(l) prohibition. The prohibition is at the system level, not the output level.

Practical Implications for EU Developers and Operators

Scenario 1: You operate a dedicated nudifier service targeting adult content platforms. Under Art.5(1)(l), this system cannot be offered to EU users from December 2027. Planning now: either geo-block EU users (with the legal risks that entails for EU residents using VPNs) or shut down the service. National law may already impose liability.

Scenario 2: You operate a general-purpose image synthesis API used by many customers. Compliance requires: usage monitoring to confirm NCII is not the primary actual use, real-person + intimate content classifiers at inference time, explicit ToS prohibition, deployer agreement clauses, and audit logging. The "primary actual use" limb means you cannot ignore how your API is actually used by customers.

Scenario 3: You operate a professional video production tool with model consent contracts. If your actors provide written, informed consent for specific synthetic intimate imagery for defined professional purposes (e.g., stunt doubles, historical recreations), the consent exception may apply. You need documented consent processes, purpose limitations, and data retention policies aligned with GDPR Art.9.

Scenario 4: You operate a CSAM detection system that generates synthetic imagery for classifier training. The professional context exception likely applies, but requires strict governance: generation limited to classifier training purposes only, access restricted to authorized researchers, no distribution of generated imagery, and documented necessity assessment.

Summary

Art.5(1)(l) — introduced by the EU AI Act Digital Omnibus — creates an absolute prohibition on AI systems whose design or primary actual use is the generation of non-consensual synthetic intimate imagery of real individuals. The prohibition:

For most providers of general-purpose image and video synthesis AI systems, the key compliance obligation is not system prohibition but primary actual use monitoring — ensuring NCII does not become the dominant use case for your system through a combination of technical controls, ToS enforcement, and deployer due diligence.

See Also